HaomingLuo/AgentFEM-DoubleNotch-Elasticity-2D

AgentFEM · Double-notch stress concentration

A tensile plate with two opposed semicircular edge notches. Study how ligament width and geometry redistribute displacement and stress.

256 independently solved parameter sets · full meshes and fields · 192 / 32 / 32 split · CC BY 4.0

Built with AgentFEM. A small, reproducible engineering dataset for surrogate learning, field prediction and numerical-method experiments.

Physical problem

Small-strain isotropic plane-stress elasticity. The entire left edge is clamped; a uniform positive x traction acts on the right edge. All other boundaries are traction free. The semicircular cutouts are centered at (L/2,0) and (L/2,H), leaving ligament H-2r.

Quadratic displacement elements on straight-sided triangular meshes; Gmsh size limits H/80 to H/40. Stress/strain are cellwise L2-projected values, not exact surface peaks.

Elastic response only: no plasticity, fatigue-life, fracture-growth, or safety approval. Clamp corners can have mesh-sensitive stresses; use the supplied central-region metric (0.3L < x < 0.7L), not an unrestricted global stress maximum.

What is included

• train.h5, validation.h5, test.h5: one group per sample ID, geometry, connectivity, point fields and cell fields. Gzip-compressed without floating-point downcasting.
• HDF5 inputs/: boundary-node masks and, for heat, the prescribed Gaussian source evaluated at mesh nodes. These are model inputs, not predicted fields or integrated nodal loads.
• Corresponding JSONL files: browsable parameters and engineering observables; they point to HDF5 groups, not serialized full fields.
• quality.json: checks, dependency versions, convergence evidence.
• baselines.json, baseline.py: held-out scalar mean/nearest-neighbor baselines with train-only normalization; no neural-model claims.
• generate.py, quality.py, load_example.py: generation, audit and minimal loading workflows.

SI units: coordinates and U in m; stress S and MISES in Pa; strain E dimensionless; temperature in K. coordinates are undeformed reference nodes. Connectivity follows the AgentFEM XDMF export order: Triangle (heat), Triangle_6 (notch), or Quadrilateral_9 (cylinder). Cell stress is projected and unsmoothed. See generated XDMF for native export semantics.

Tensor components are flattened row-major: 2×2 for plane stress (x,y), 3×3 for axisymmetry (r,theta,z). Displacement U uses padded three-component storage; active components are (x,y) for the plate and (r,z) for the cylinder. The HDF5 files contain mesh connectivity, not an assumed common image grid.

Parameters, in generator order: length_m, height_m, notch_radius_m, young_pa, poisson, traction_pa.

Observed sample ranges (exact values are in JSONL; intended sampling bounds are in generate.py):

Parameter	Minimum	Maximum
length_m	0.12006	0.199827
height_m	0.0401169	0.0699237
notch_radius_m	0.00503543	0.0202624
young_pa	7.01094e+10	2.09986e+11
poisson	0.220298	0.33993
traction_pa	1.00666e+07	4.98825e+07

Quality evidence

Eight mesh-refinement audits: maximum relative change in central_cell_mises_max_pa = 5.49%. This is an observable convergence check, not a bound on full-field error.

All published field arrays are finite; all parameter sets and geometry/source-layout parameter tuples are unique. Splits use one fixed random permutation after scrambled Sobol sampling (seed 20260923; split seed 20260924). All splits are in-distribution; no claim of unseen-family generalization. One row is one independently solved case, not a frame or a rescaled duplicate.

Generic continuous material ranges are synthetic, not calibrated material grades. Dataset generation does not imply experimental validation. These are original parameterized examples, not certified NAFEMS benchmarks or copies of commercial solver datasets.

Load and reproduce

python load_example.py --repo HaomingLuo/AgentFEM-DoubleNotch-Elasticity-2D --split test
# In the pinned AgentFEM/FEniCSx environment:
python generate.py notch --count 256 --refinement 2

Loading requires only h5py, numpy and huggingface_hub; FEniCSx is needed only to regenerate FEM results. Generation writes a local generated/ directory and uses one MPI rank. The heat manufactured-solution check is available via python generate.py heat --validate.

Attribution

Haoming Luo, Double-notch stress concentration — AgentFEM parametric FEM dataset, 2026. Please cite this repository URL and the commit revision used. Data: CC BY 4.0. Example code: Apache-2.0.